Audio Adversarial Examples: Targeted Attacks on Speech-to-Text

Nicholas Carlini, David A. Wagner

Abstract

We construct targeted audio adversarial examples on automatic speech recognition. Given any audio waveform, we can produce another that is over 99.9% similar, but transcribes as any phrase we choose (recognizing up to 50 characters per second of audio). We apply our white-box iterative optimization-based attack to Mozilla's implementation DeepSpeech end-to-end, and show it has a 100% success rate. The feasibility of this attack introduces a new domain to study adversarial examples.

Introduction

While adversarial examples—inputs designed to fool machine learning models—have been extensively studied in computer vision, the audio domain has remained more resilient. Previous research into audio attacks was often limited to untargeted failures or required synthesizing entirely new, unnatural sounds.

The complexity of Automatic Speech Recognition (ASR) systems, which involve preprocessing (MFCC conversion), temporal modeling (RNNs), and decoding (CTC), presents a significant challenge for adversaries. Researchers Nicholas Carlini and David Wagner have now demonstrated that these systems are just as vulnerable as image recognition models. They show that by applying white-box iterative optimization, any audio source can be modified with a nearly inaudible layer of noise to force a specific, targeted transcription with a 100% success rate.

Methodology

The attack is framed as an optimization problem targeting Mozilla’s DeepSpeech implementation. The goal is to modify an existing audio file $x$ into $x + \delta$ such that it sounds nearly identical to $x$ but transcribes exactly as a chosen target phrase $y$.

Overview of the targeted audio adversarial attack process.

Key technical components include:

  • Differentiating through the pipeline: Gradients are calculated through the entire ASR process, from the final transcription loss back to the raw waveform.
  • Two-Step Optimization: The process first identifies a "path" (alignment of characters to audio frames) using CTC loss. Then, a custom loss function minimizes the perturbation $\delta$ to ensure it remains as quiet as possible.
  • Distortion Metric: Loudness is measured in Decibels (dB), with the objective of keeping the perturbation significantly quieter than the original signal.

Evaluation and Results

The attack was evaluated using the Mozilla Common Voice dataset and clips of classical music. For each of the first 100 test samples in Common Voice, 10 different random target transcriptions were generated.

The results were definitive: a 100% success rate across all test cases. The mean distortion was -38dB relative to the original signal—roughly the difference between a person talking and ambient noise in a quiet room.

Comparison of the original (blue) and adversarial (orange) waveforms, demonstrating high similarity.

Discussion and Insights

The research highlights several unique aspects of the audio domain:

  • Non-Linearity: Unlike image recognition, audio processing is highly non-linear due to MFCC preprocessing and deep LSTMs, making iterative optimization essential.
  • High Information Density: The attack can push the model to its theoretical limit of 50 characters per second, allowing short clips to be transcribed into long sentences.
  • Versatility: The attack does not require a human voice as a starting point; even classical music can be turned into a specific voice command.

Limitations and Future Work

Currently, these adversarial examples are "brittle" and often fail when played over-the-air due to physical distortions. Additionally, while the attack can be optimized against MP3 compression, doing so increases the required loudness of the perturbation. These findings underscore the need for the research community to develop defenses specifically tailored to the unique non-linearities of audio processing.

Summary updated

Abstract

We construct targeted audio adversarial examples on automatic speech recognition. Given any audio waveform, we can produce another that is over 99.9% similar, but transcribes as any phrase we choose (recognizing up to 50 characters per second of audio). We apply our white-box iterative optimization-based attack to Mozilla's implementation DeepSpeech end-to-end, and show it has a 100% success rate. The feasibility of this attack introduce a new domain to study adversarial examples.
BibTeX
@inproceedings{carlini2018audio,
  author = {Nicholas Carlini and David A. Wagner},
  title = {{Audio Adversarial Examples: Targeted Attacks on Speech-to-Text}},
  booktitle = {IEEE Symposium on Security and Privacy Workshops},
  year = {2018},
  url = {https://openreview.net/forum?id=lH9FAMDjcCV},
}
Figure